台达电氢能主管论坛阐述AI资料中心燃料电池+电池分层供电与800V直流架构，属愿景性发言、无订单等增量事实

Artificial intelligence data-center expansion is intensifying a global race for electricity, as cloud providers increasingly prioritize how quickly power can be secured over cost and decarbonization. As grids struggle to keep pace with computing growth, companies are considering on-site generation, batteries, and direct-current systems to speed deployment while managing emissions and efficiency.

The AI race has increasingly become a contest between time-to-market and time-to-power, according to Charles Tsai, general manager of Delta Electronics' Hydrogen Energy Department. Speaking at a forum, Tsai said limited electricity availability is becoming the biggest obstacle for AI developers.
Application reviews and grid-connection work can take a long time, while public-grid expansion has struggled to keep pace with computing capacity. Discussions with customers in the UK and international cloud service providers indicated that power-evaluation priorities had shifted, with time-to-power ranking first, followed by cost and decarbonization.
Tsai said conventional large-scale gas-fired power plants typically take six to 10 years to build, while equipment delivery schedules for some international heavy-industry manufacturers have also reached 10 years. By comparison, onsite systems such as solid oxide fuel cells can require about two years from order placement to delivery.
Fuel cells, however, cannot respond quickly enough to the millisecond- or microsecond-level power changes that AI server motherboards require. Tsai proposed a layered architecture that combines fuel cells with high-power batteries to support real-time adjustment, capacity optimization, and peak shaving and valley filling.
He also said fuel-cell exhaust can contain carbon dioxide concentrations of up to 60%, making the emissions easier to capture through carbon capture and storage than those from conventional natural-gas or coal-fired generation. Recovering high-temperature waste heat for warm water could further improve efficiency and increase economic benefits.
For next-generation AI data centers, Tsai proposed an 800V direct-current architecture that would send electricity from generation equipment directly to servers. By reducing the number of transformers and AC-to-DC conversion facilities, the approach could cut power-infrastructure costs by as much as one-third and help address Taiwan Power's grid capacity limits and lengthy approval processes.